It isn't if you actually listened to how the solutions work and differ.
Wait, I thought PS5 Pro had dedicated Tensor cores for PSSR?
No, the alterations to the WGPs mean that the feature set can't be provided by the PS5 even if they scaled everything right down. The 300TOPs figure comes from them doing 3x calculations of an inference per clock IIRC and then doing INT8/INT4 quantised CNNs, which again would require the PS5 OG to do as INT16 at orders less efficient, and IIRC there's also a ASIC for doing the non-inference scaling.
So, does PS5 Pro have dedicated hardware (Tensor/Matrix cores)?
No, but the stacked caches - which was the major customisation to WGPs IIRC - provides useable memory bandwidth to the light CNNs way beyond the dedicated tensors for such light workloads. Cerny's recent discussion of the Pro hardware is another masterclass and well worth 30mins.
The CUs are modified to process additional instructions needed for ML inference.
Hmmm, interesting and maybe a bit disappointing...
Matrix Cores since UDNA, rest answers you got above
Blackwell also supports FP4, so if DLSS4 utilizes less accuracy, it won't be possible.
Yes, I corrected my post after that. I was wondering the performance cost and how that will impact lower tier and older architectures. 5% on what is the question? 5% on a 4090? 5% on a 2060?It's confirmed that it does
I'm sure you saw this chart?
The Nvidia tech marketing manager on twitter posted the difference of performance loss between the two models, CNN & Transformer
" Yes, it is a little heavier, but the loss must be small, less than 5%"
If it runs on Turing's old tensor cores which had no concurrency with shading pipeline, it's not a problem for any architectures afterwards.
The best of Turing's tensor core performances was 89 TOPs. Ampere bitchslaps that, even the Tegra Orin architecture and the estimates of Switch with 48 Tensor cores would hover at ~100 TOPS
The thing is the new model looks so good that you can effectively go into maybe even lower tiers of DLSS and gain performance. If you were fine with DLSS Quality, you can probably go down to DLSS Balanced if not performance and see equivalent results.
We'll see soon enough, day 0 of Blackwell launch we can inject the new DLSS 4 into all DLSS 2 titles.
Yeah. As long as the game already supports any form of DLSS, DLSS 4 and it's new model will be available to use on that game.
That's incredible! Can't wait to test how it does on Final Fantasy XV, which was among the first titles to support DLSS back in 2018.Yeah. As long as the game already supports any form of DLSS, DLSS 4 and it's new model will be available to use on that game.
Won’t work, FF 15 still uses DLSS 1 which is something completely different. I think it’s actually the only game left that still has it.
The stacked cache bandwidth discussion tells me it has already won out because when 1ms is the fraction of utilisation the silicon gets in a 8ms, 16ms or 33ms frame and to divide and conquer the bandwidth requirement, and handle the latency constraints the CNNs need to tile across WGP caches it just sounds counter productive in terms of design to be building more powerful, more expensive and more wasteful ML accelerators with more wasteful highspeed interconnects when you can generalise the functionality with WGPs and a more complex software interface.
He’s most likely the person who will create Skynet…
I'm wondering why Sony cannot do the same with PSSR, but instead game devs have to patch games for the latest PSSR version.Yeah. As long as the game already supports any form of DLSS, DLSS 4 and it's new model will be available to use on that game.
Interesting explanation.
Question: Are GPU architectures and Direct3D evolving toward a design where the distinction between vertex and pixel shaders essentially goes away?—davesalvator
David Kirk: For hardware architecture, I think that's an implementation detail, not a feature.
For sure, the distinction between the programming models and instruction sets of vertex shaders and pixel shaders should go away. It would be soooo nice for developers to be able to program to a single instruction set for both.
As to whether the architectures for vertex and pixel processors should be the same, it's a good question, and time will tell the answer. It's not clear to me that an architecture for a good, efficient, and fast vertex shader is the same as the architecture for a good and fast pixel shader. A pixel shader would need far, far more texture math performance and read bandwidth than an optimized vertex shader. So, if you used that pixel shader to do vertex shading, most of the hardware would be idle, most of the time. Which is better—a lean and mean optimized vertex shader and a lean and mean optimized pixel shader or two less-efficient hybrid shaders? There is an old saying: "Jack of all trades, master of none."
~
Reader Question: Will we see non-uniform rasterization, streaming ray-casts, or equivalent features to enable the kind of graphics we really want—real-time, dynamic lighting with a large number of lights?—Jason_Watkins
David Kirk: Yes.
Over time, everything that has been hardwired and fixed-function will become general-purpose. This will enable much more variety in graphics algorithms and ultimately, much more realism.
The good news, for my job security, is that graphics is still ultimately very very hard. Tracing streaming rays in all directions, reflecting between all of the objects, lights, and fog molecules in parallel is extremely hard. Nature does it . . . in real time! However, nature does it by using all of the atoms in the universe to complete the computation. That is not available to us in a modern GPU.
Graphics will continue to be a collection of clever tricks, to do just enough work to calculate visibility, lighting, shadows, and even motion and physics, without resorting to brutishly calculating every detail. So, I think that there's a great future both in more powerful and flexible GPUs as well as ever more clever graphics algorithms and approaches.
~
Question: With all of the pressing towards more powerful graphics cards to handle features such as FSAA and anisotropic filtering, why do we still use inefficient, "fake" methods to achieve these effects?—thalyn
David Kirk: Looking at my answer from the question about ray-casting and lighting effects, graphics is all "fake" methods. The trick is to perform a clever fake and not get caught! All graphics algorithms do less work than the real physical universe but attempt to produce a realistic simulation.
Is that because DLSS1 required training in the cloud? Or something like that, IIRC.
It won't work, RTX 40 series lack certain circuitry (Flip Metering):
The AMD RDNA4 approach has already lost, UDNA is switching to dedicated ML units (Matrix Accelerators) like CDNA2/3 has.
Yeah me too, I think that sucks, and same with games locked to 30fps. I guess the games just weren't ready for the new tech, should be smooth sailing for any new games.
Were did you get that from? Afaik, only thing that affects latency is frame gen, not the AI scaling.
Checked on the web and seems like it will release with the 5xxx series.
Eventually IMO, but I think Nvidia will fight it with marketing for a coupe of years. But the gains for PS6 and future RDNA low power devices to maximise the versatility of pricey cache modules and scale ML performance proportionally with price/size seem like a nice advantage to offer more bang per buck. The MFG latency stuck at a pointless 33ms for 90% fake frames at +200fps feels like a marketing stop gap from Nvidia that they'll ditch if they were competing with AMD with at real 90fps with FSR4 at DLSS2 image quality + ray reconstruction and latency of 11ms.
Well I watched it all and understood about 5% of it, sober, so yeah I can agree on that. But he’s extremely good on stage, maybe too good because he doesn’t let anybody else up there.
There's no elephant in the room. People's eyesight is not to blame like you think, the DLSS image simply looks better.
Lol. No. But whatever.
AI DLSS, the 1.0 version, has gigazillion of pixels and had been trained in datacenters on zillion of GPUs.
There's no elephant in the room. People's eyesight is not to blame like you think, the DLSS image simply looks better.
You still think that DLSS simply upscales lower resolution, and that's the problem. You can't understand how it's possible to render more detail from lower resolution image.In reality DLSS doesnt upscale resolution. DLSS image reconstruction (that's how Nvidia engineers call it) reduce internal resolution in the first step (to improve framerate), but it will rebuild resolution in the final image based on data from multiple previous frames.
4K native frame has 8 million pixels, but DLSS (quality) will use multiple 1440p 3.6 million pixel frames to generate the final image. DLSS has data consisting from well over 20 million of pixels, and that's why DLSS image can look better than 8 million pixels frame (4K TAA).
Shadow Of The Tomb Raider used version 2.0, and DLSS image was already better than native 4K. You can clearly see more detail in the DLSS image in the distant buildings and textures.
DLSS technology has improved further since then and now the latest 3.8.1 DLSS can outperform native TAA even in performance mode.
TAA Native
DLSS 3.8.1 Quality
DLSS 3.8.1 Performance
TAA native
DLSS 3.8.1 Quality
DLSS 3.8.1 Performance
TAA has a little bit more sharpening mask, compared to DLSS default settings, but the DLSS image have clearly more detail in the static image and it's also not blurred during motion (the difference is huge in this particular game). With reshade filters on top of that, the DLSS image is perfectly sharp, but for some strange reason reshade filters are not saved on screenshots when games use the Vulkan API.
As anyone can see, even DLSS performance can deliver better image quality than native TAA, and if you combine "DLSS performance" + "DLDSR2.25", the TAA looks like an upscaled image in comparison (and you still get big performance boost even with DLDSR!!). If you only own RTX card it simply doesnt make sense to play at TAA native, because you will get not only worse image quality, but also performance. DLSS image reconstruction will soon be upgraded with a new 'transformer' model, and perhaps even DLSS ultra performance will be able to beat native TAA (will see).
We have to wait until FSR 4 is widespread before ML based upscaling magically stops becoming "evil".
It's funny how he ignored my comparison, which clearly showed that the DLSS image had more detail compared to the TAA native. This guy ignores reality, and would rather believe that people have eyesight problems than admit that DLSS offers better results.
The AMD guys can't use DLSS and that's probably why they pretend TAA looks better, but you're right, once FSR supports ML on radeon cards and looks comparable to DLSS they'll finally start to see reality.
There's no elephant in the room. People's eyesight is not to blame like you think, the DLSS image simply looks better.
You still think that DLSS simply upscales lower resolution, and that's the problem. You can't understand how it's possible to render more detail from lower resolution image.In reality DLSS doesnt upscale resolution. DLSS image reconstruction (that's how Nvidia engineers call it) reduce internal resolution in the first step (to improve framerate), but it will rebuild resolution in the final image based on data from multiple previous frames.
4K native frame has 8 million pixels, but DLSS (quality) will use multiple 1440p 3.6 million pixel frames to generate the final image. DLSS has data consisting from well over 20 million of pixels, and that's why DLSS image can look better than 8 million pixels frame (4K TAA).
Shadow Of The Tomb Raider used version 2.0, and DLSS image was already better than native 4K. You can clearly see more detail in the DLSS image in the distant buildings and textures.
DLSS technology has improved further since then and now the latest 3.8.1 DLSS can outperform native TAA even in performance mode.
TAA Native
DLSS 3.8.1 Quality
DLSS 3.8.1 Performance
TAA native
DLSS 3.8.1 Quality
DLSS 3.8.1 Performance
TAA has a little bit more sharpening mask, compared to DLSS default settings, but the DLSS image have clearly more detail in the static image and it's also not blurred during motion (the difference is huge in this particular game). With reshade filters on top of that, the DLSS image is perfectly sharp, but for some strange reason reshade filters are not saved on screenshots when games use the Vulkan API.
As anyone can see, even DLSS performance can deliver better image quality than native TAA, and if you combine "DLSS performance" + "DLDSR2.25", the TAA looks like an upscaled image in comparison (and you still get big performance boost even with DLDSR!!). If you only own RTX card it simply doesnt make sense to play at TAA native, because you will get not only worse image quality, but also performance. DLSS image reconstruction will soon be upgraded with a new 'transformer' model, and perhaps even DLSS ultra performance will be able to beat native TAA (will see).
